Semi-supervised NMF with time-frequency annotations for singlechannel source separation

We formulate a novel extension of nonnegative matrix fac-torization (NMF) to take into account partial information on source-specific activity in the spectrogram. This infor-mation comes in the form of masking coefficients, such as those found in an ideal binary mask. We show that state-of-the-art results in source separation may be achieved with only a limited amount of correct annotation, and further-more our algorithm is robust to incorrect annotations. Since in practice ideal annotations are not observed, we propose several supervision scenarios to estimate the ideal mask-ing coefficients. First, manual annotations by a trained user on a dedicated graphical user interface are shown to provide satisfactory performance although they are prone to errors. Second, we investigate simple learning strate-gies to predict the Wiener coefficients based on local in-formation around a given time-frequency bin of the spec-trogram. Results on single-channel source separation show that time-frequency annotations allow to disambiguate the source separation problem, and learned annotations open the way for a completely unsupervised learning procedure for source separation with no human intervention. 1